Top Banner
A Simple Derivation of Hadamard’s Variational Formula Nuha Loling OTHMAN , 1 Takashi SUZUKI ( ), 2 Takuya TSUCHIYA ( ), 3 $\mathbb{R}^{n}$ Poisson Green 100 Hadamard Green Green Hadamard Hadamard Green Garabedian-Schiffer 1 Introduction Let $\mathbb{R}^{n}$ be n-dimensional Euclidean space $(n\geq 2)$ and $\Omega\subset \mathbb{R}^{n}$ be a bounded domain. For a given function $f$ , we consider Poisson’s equation $-\triangle u=f$ on $\Omega$ , $u=0$ on $\partial\Omega$ . The Green function $G(x, y)$ is a function which provides the solution $u$ of the Poisson equation by $u(x)= \int_{\Omega}G(x, y)f(y)dy$ . If the domain $\Omega$ is modified, then the Green function $G(x, y)$ would vary. Hadamard considered how $G(x, y)$ would vary and computed the first variation $\delta G(x, y)$ with respect to domain per- turbation [3]. His result is now called Hadamard’s variational formula. Hadamard showed his formula under the assumption that $\partial\Omega$ and the perturbation are analytic. Later, Garabedian and Schiffer gave a simpler and more rigorous proof of Hadamard‘s variational formula under the assumption that $\partial\Omega$ and the perturbation are of $C^{2}$ class (see [1]). Further, they obtained Hadamard’s second variational formula [2], [4]. The main aim of this paper is to reconsider Hadamard’s variational formula. In particular, we develop a methodology which provides us a much clearer understanding of Hadamard’s variational formula. As a result, we obtain a very simple proof of Hadamard’s variational formula (see Section 3.1). We also obtain Hadamard’s second variational formula which is an extension of Grabedian-Schiffer’s formula (Theorem 3.3). Here, we briefly summarize the notation which we use in this paper. We denote the Euclidean inner product by $x\cdot y$ or $(x, y)_{\mathbb{R}^{n}}$ for $x,$ $y\in \mathbb{R}^{n}$ . When we do not specify, all vectors in $\mathbb{R}^{n}$ are regarded as column vectors. Transposing of vectors and matrices are denoted by $(\cdot)^{T}$ . Let $f(x)$ be a smooth function defined in a domain of $\mathbb{R}^{n}$ . The gradient of $f$ is denoted by $\nabla f(x):=(\frac{\partial f}{\partial x_{1}}(x),$ $\cdots,$ $\frac{\partial f}{\partial x_{n}}(x))$ . lGraduate School of Engineering Science, Osaka University, othmanQsigmath. es. osaka-u. ac. jp 2Graduate School of Engineering Science, Osaka University, suzukiQsigmath. es. osaka-u. ac. jp 3Graduate School of Science and Engineering, Ehime University, tsuchiyaQmath. sci. ehime-u. ac. jp 1733 2011 127-141 127
15

Hadamard’s of Akyodo/kokyuroku/contents/pdf/...Hadamard’s second variational formula [2],[4]. Themain aim thisof is paper to reconsider Hadamard’s variational formula. In particular,

Feb 09, 2021

Download

Documents

dariahiddleston
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
  • A Simple Derivation ofHadamard’s Variational Formula

    Nuha Loling OTHMAN, 1 Takashi SUZUKI (鈴木 貴 ), 2 Takuya TSUCHIYA (土屋 卓也), 3

    ここでは、 $\mathbb{R}^{n}$ 内の有界領域上での Poisson 問題と、 その問題の Green 関数を考える。約 100年前、Hadamard は、領域の境界が摂動を受けた際に Green 関数がどのような影響を受けるかという問題を考え、領域の摂動に対する Green 関数の第一変分を求めた。 それは、 現在 Hadamard の変分公式と呼ばれている。 この論文では、 Hadamard の変分公式の別証明を与える。 さらに、領域の摂動に対する Green 関数の第二変分も計算することができた。我々の公式は、 Garabedian-Schifferの公式の拡張になっている。

    1 Introduction

    Let $\mathbb{R}^{n}$ be n-dimensional Euclidean space $(n\geq 2)$ and $\Omega\subset \mathbb{R}^{n}$ be a bounded domain. For agiven function $f$ , we consider Poisson’s equation

    $-\triangle u=f$ on $\Omega$ , $u=0$ on $\partial\Omega$ .

    The Green function $G(x, y)$ is a function which provides the solution $u$ of the Poisson equationby

    $u(x)= \int_{\Omega}G(x, y)f(y)dy$ .

    If the domain $\Omega$ is modified, then the Green function $G(x, y)$ would vary. Hadamard consideredhow $G(x, y)$ would vary and computed the first variation $\delta G(x, y)$ with respect to domain per-turbation [3]. His result is now called Hadamard’s variational formula. Hadamard showedhis formula under the assumption that $\partial\Omega$ and the perturbation are analytic. Later, Garabedianand Schiffer gave a simpler and more rigorous proof of Hadamard‘s variational formula underthe assumption that $\partial\Omega$ and the perturbation are of $C^{2}$ class (see [1]). Further, they obtainedHadamard’s second variational formula [2], [4]. The main aim of this paper is to reconsiderHadamard’s variational formula. In particular, we develop a methodology which provides us amuch clearer understanding of Hadamard’s variational formula. As a result, we obtain a verysimple proof of Hadamard’s variational formula (see Section 3.1). We also obtain Hadamard’ssecond variational formula which is an extension of Grabedian-Schiffer’s formula (Theorem 3.3).

    Here, we briefly summarize the notation which we use in this paper. We denote the Euclideaninner product by $x\cdot y$ or $(x, y)_{\mathbb{R}^{n}}$ for $x,$ $y\in \mathbb{R}^{n}$ . When we do not specify, all vectors in $\mathbb{R}^{n}$ areregarded as column vectors. Transposing of vectors and matrices are denoted by $(\cdot)^{T}$ . Let $f(x)$be a smooth function defined in a domain of $\mathbb{R}^{n}$ . The gradient of $f$ is denoted by

    $\nabla f(x):=(\frac{\partial f}{\partial x_{1}}(x),$ $\cdots,$$\frac{\partial f}{\partial x_{n}}(x))$ .

    lGraduate School of Engineering Science, Osaka University, othmanQsigmath. es. osaka-u. ac. jp2Graduate School of Engineering Science, Osaka University, suzukiQsigmath. es. osaka-u. ac. jp3Graduate School of Science and Engineering, Ehime University, tsuchiyaQmath. sci. ehime-u. ac. jp

    数理解析研究所講究録第 1733巻 2011年 127-141 127

  • When we need specify the variable of a gradient, we denote such as $\nabla_{x}f(x),$ $\nabla_{x}*f(x^{*})$ . Weregard gradients of functions as row vectors. Hence, for a vector field $F(x),$ $\nabla F(x)$ is the Jacobimatrix $DF(x)$ . Let $\Omega\subset \mathbb{R}^{n}$ be a domain in $\mathbb{R}^{n}$ . We denote by $L^{2}(\Omega),$ $H^{1}(\Omega),$ $H^{s}(\partial\Omega)$ the usualLebesgue and Sobolev spaces. The inner product of $L^{2}(\Omega)$ is denoted by

    $(u, v)_{\Omega}$ $:= \int_{\Omega}$ uvdx, $u,$ $v\in L^{2}(\Omega)$

    On a point $x\in\partial\Omega$ , we denote the unit outer normal vector of $\partial\Omega$ by $\nu=\nu(x)$ . For asubset $\Gamma\subset\partial\Omega$ , we denote the duality pair of $H^{-1/2}(\Gamma)$ and $H^{1/2}(\Gamma)$ by $\langle\varphi,$ $v\rangle_{\Gamma},$ $\varphi\in H^{-1/2}(\Gamma)$ ,$v\in H^{1/2}(\Gamma)$ .

    2 Basic Definitions

    Let $\Omega\subset \mathbb{R}^{n}$ be a bounded Lipschitz domain and $\tilde{\Omega}$ be a sufficiently larger domain which satisfies$\overline{\Omega}\subset$ int $\tilde{\Omega}$ . For a parameter $t\geq 0$ , we define transformation $\mathcal{T}_{t}$ : $\Omegaarrow \mathcal{T}_{t}(\Omega)\subset \mathbb{R}^{n}$ of $\Omega$with respect $t$ in the following way. Let a $C^{0,1}$ -class vector field $S(x)$ be given. We supposethat suppS $\subset\tilde{\Omega}$ . Then, a transformation $\mathcal{T}_{t}(x)$ on fi is defined as a solution of the ordinarydifferential equation

    (2.1) $\frac{d}{dt}\mathcal{T}_{t}(x)=S(\mathcal{T}_{t}(x))$ , $\mathcal{T}_{0}(x)=x$ .

    That is, for each $x\in\tilde{\Omega},$ $\mathcal{T}_{t}(x)$ is the integral curve generated by (2.1). This $\mathcal{T}_{t}(x)$ satisfies thefollowing properties:

    $\bullet$ For any $x\in\Omega,$ $T_{0}(x)=x$ .. For a sufficiently small $t,$ $\Omega_{t}$ $:=\mathcal{T}_{t}(\Omega)\subset\tilde{\Omega}$ .$\bullet$ $\mathcal{T}_{t}$ is a diffeomorphism for a sufficiently small $t\geq 0$ .

    $\bullet$ $\mathcal{T}_{t}$ is smooth with respect to $t$ .

    From the definition (2.1) we have $S(x)=\frac{\partial}{\partial t}\mathcal{T}_{t}(x)|_{t=0}$ . Moreover, we define

    $T(x):=\frac{\partial^{2}}{\partial t^{2}}\mathcal{T}_{t}(x)|_{t=0}$ .

    Then, the transformation has the Taylor expansion

    $\mathcal{T}_{t}(x)=x+tS(x)+\frac{1}{2}t^{2}T(x)+o(t^{2})$

    with respect to $t$ . Here, $o(t^{2})$ denote a quantity which would be expressed by $t^{2}\omega(x, t)$ , where$\omega(x, t)$ is a function which converges uniformly (with respect to x) to $0$ as $tarrow+O$ . In the sequel,notations such as $o(t),$ $o(t^{2})$ are understood in this way. Let $DS(x)$ be the Jacobi matrix of S.From (2.1), we have

    $\frac{d^{2}}{dt^{2}}\mathcal{T}_{t}(x)=\frac{d}{dt}S(\mathcal{T}_{t}(x))=DS(\mathcal{T}_{t}(x))\frac{d}{dt}\mathcal{T}_{t}(x)=DS(\mathcal{T}_{t}(x))S(\mathcal{T}_{t}(x))$ ,

    which implies

    (2.2) $T(x)=(DS(x))S(x)$ .

    128

  • Let a function $\varphi$ be defined on $\tilde{\Omega}$ and $\varphi\in H^{2}(\tilde{\Omega})$ . Suppose that a function $u=u(x, t)\in$$H^{1}(\Omega_{t})$ is a solution of the boundary value problem

    (2.3) $\{\begin{array}{ll}\triangle v(\cdot, t)=0 in \Omega_{t},u(\cdot, t)=\varphi on \partial\Omega_{t}.\end{array}$

    Here, $\triangle$ $:=\partial^{2}/\partial x_{1}^{2}+\cdots+\partial^{2}/\partial x_{n}^{2}$ is the usual Laplacian with respect to $x=(x_{1}, \cdots, x_{n})^{T}$ . In thissection, we investigate differentiations of quantities which depend on $u(x, t)$ . Such variations ofquantities with respect to domain perturbation are called Hadamard $s$ variation. To computeHadamard’s variation it is important to know Lagrangian derivative 4 $\dot{u}_{\mathcal{L}},$ $ii_{\mathcal{L}}$ and Eulerianderivative 5 $\dot{u}\mathcal{E},\ddot{u}\mathcal{E}$ , defined by, for $x\in\Omega$ ,

    $\dot{u}_{\mathcal{L}}(x):=\frac{d}{dt}(u(\mathcal{T}_{t}(x), t))|_{t=0}$ $\ddot{u}_{\mathcal{L}}(x):=\frac{d^{2}}{dt^{2}}(u(\mathcal{T}_{t}(x), t))|_{t=0}$ ,

    $\dot{u}_{\mathcal{E}}(x):=\frac{\partial}{\partial t}u(x, t)|_{t=0}$ , $\ddot{u}_{\mathcal{E}}(x):=\frac{\partial^{2}}{\partial t^{2}}u(x, t)|_{t=0}$ .

    For a function $f(x, t)$ , let

    $\mathcal{H}_{x}f=\mathcal{H}_{x}f(x, t):=(\frac{\partial^{2}f(x,t)}{\partial x_{i}\partial x_{j}})_{i,j=1,\cdots,n}$

    be the Hesse matrix. We use the same notation $\mathcal{H}_{x}f$ for the second order tensor $\mathcal{H}_{x}f$ : $\mathbb{R}^{n}\cross \mathbb{R}^{n}arrow$$\mathbb{R}$ defined by $\mathcal{H}_{x}f(X, Y)$ $:=((\mathcal{H}_{x}f)X, Y)_{\mathbb{R}^{n}}$ for $X,$ $Y\in \mathbb{R}^{n}$ . In particular, in the case of $X=Y$ ,we denote as $\mathcal{H}_{x}f(X, X)=\mathcal{H}_{x}f\cdot(X)^{2}$ . A straightforward computation yields

    (2.4) $\frac{d}{dt}(u(\mathcal{T}_{t}(x), t))=\frac{\partial}{\partial t}u(\mathcal{T}_{t}(x), t)+\nabla u(\mathcal{T}_{t}(x), t)\cdot(\frac{\partial}{\partial t}\mathcal{T}_{t}(x))$ ,

    (2.5) $\frac{d^{2}}{dt^{2}}(u(\mathcal{T}_{t}(x), t))=\frac{\partial^{2}}{\partial t^{2}}u(\mathcal{T}_{t}(x), t)+2\nabla(\frac{\partial}{\partial t}u(\mathcal{T}_{t}(x), t))\cdot(\frac{\partial}{\partial t}\mathcal{T}_{t}(x))$

    $+ \nabla u(\mathcal{T}_{t}(x), t)\cdot(\frac{\partial^{2}}{\partial t^{2}}\mathcal{T}_{t}(x))+\mathcal{H}_{x}u(\mathcal{T}_{t}(x), t)\cdot(\frac{\partial}{\partial t}\mathcal{T}_{t}(x))^{2}$ .

    2.1 Eulerian Derivatives $\dot{u}_{\mathcal{E}},\ddot{u}_{\mathcal{E}}$

    In this subsection, we check properties which Eulerian derivatives $\dot{u}_{\mathcal{E}},\ddot{u}_{\mathcal{E}}$ should satisfy. At aninner point $x\in\Omega$ we have $\triangle u(\cdot, t)=0$ for any $t$ . Hence,

    $\triangle\dot{u}_{\mathcal{E}}=0$ , $\triangle\ddot{u}_{\mathcal{E}}=0$ in $\Omega$ .

    On the boundary $\partial\Omega$ we have $u(\mathcal{T}_{t}(x), t)=\varphi(\mathcal{T}_{t}(x))$ . Differentiating the both side and letting$tarrow+0$ , we see $\dot{u}_{\mathcal{E}}+$ S. $\nabla u=$ S. $\nabla\varphi$ . Therefore, we find that the Eulerian derivative $\dot{u}s$ is asolution of the following boundary value problem:

    (2.6) $\triangle\dot{u}\mathcal{E}=0$ in $\Omega$ , $\dot{u}\mathcal{E}=S$ . $(\nabla\varphi-\nabla u)$ on $\partial\Omega$ .

    In the same manner, we conclude that $\ddot{u}_{\mathcal{E}}$ is a solution of the boundary value problem$\triangle\ddot{u}_{\mathcal{E}}=0$ in $\Omega$ ,

    (2.7)$\ddot{u}_{\mathcal{E}}=-2S$ . $\nabla\dot{u}_{\mathcal{E}}+T$ . $(\nabla\varphi-\nabla u)+(\mathcal{H}_{x}\varphi-\mathcal{H}_{x}u)\cdot(S)^{2}$ on $\partial\Omega$ .

    4It is also called material derivative or covariant demvative.5This ib a usual partial derivative with respect to $t$ which is also called shape derivative.

    129

  • 2.2 Lagrangian Derivatives $\dot{u}_{\mathcal{L}},\ddot{u}_{\mathcal{L}}$

    In this subsection, we check properties which Lagrangian derivatives $\dot{u}_{\mathcal{L}},\ddot{u}_{C}$ should satisfy. Here,variable on $\Omega_{t}$ is denoted as $x^{*}=\mathcal{T}_{t}(x)$ . A function $f(x^{*})$ defined on $\Omega_{t}$ is pulled back by $\mathcal{T}_{t}$ toa function $f(x)$ on $\Omega$ as

    $f(x):=f(T_{t}(x))$ .

    Note that we have

    $\nabla_{x’}f=(\frac{\partial f}{\partial x_{1}},$$\cdots,$

    $\frac{\partial f}{\partial x_{n}})(\begin{array}{lll}\frac{\partial x\iota}{\partial x_{1}^{*}} \frac{\partial x1}{\partial x_{7l}}| . |\frac{\partial}{\partial}xAx_{1}^{*} \cdots \frac{\partial}{\partial}x_{A}x_{n}^{*}\end{array})=( \nabla_{x}f)(D\mathcal{T}_{t}^{-1})$ ,

    where $DT_{t}^{-1}$ is the Jacobi matrix of $\mathcal{T}_{t}^{-1}$ .The weak form of the boundary value problem (2.3) is

    (2.8) $\{\begin{array}{l}(\nabla u(\cdot, t), \nabla\tilde{v})_{\Omega_{t}}=0, \forall\tilde{v}\in H_{0}^{1}(\Omega_{t}),u(\cdot, t)=\varphi on \partial\Omega_{t}.\end{array}$

    Using the transformation $\mathcal{T}_{t}$ , we pull back the problem (2.8) to a problem defined on $\Omega$ . Notethat

    $\overline{\tau)}\in H_{0}^{1}(\Omega_{t})\Leftrightarrow v:=\tilde{v}\circ \mathcal{T}_{t}\in H_{0}^{1}(\Omega)$.

    Then, setting $u_{t}(x);=u(\mathcal{T}_{t}(x), t)$ , we see that

    $( \nabla u(\cdot, t), \nabla\tilde{v})_{\Omega_{f}}=\int_{\Omega}(\det DT_{t})(\nabla u_{t}(D\mathcal{T}_{t}^{-1}0\mathcal{T}_{t})(D\mathcal{T}_{t}^{-1}0\mathcal{T}_{t})^{T})\cdot\nabla vdx$

    $=(A(t)\nabla u_{t}, \nabla\iota))_{\Omega}$ , $\forall v\in H_{0}^{1}(\Omega)$ ,

    where$A(t)$ $:=(\det D\mathcal{T}_{t})(D\mathcal{T}_{t}^{-1}0\mathcal{T}_{t})(D\mathcal{T}_{t}^{-1}0\mathcal{T}_{t})^{T}$ .

    That is, the boundary value problem (2.8) on $\Omega_{t}$ is pulled back to the boundary value problem

    (2.9) $\{\begin{array}{l}(A(t)\nabla u_{t}, \nabla v)_{\Omega}=0, \forall v\in H_{0}^{1}(\Omega),u_{t}=\varphi oT_{t} on \partial\Omega\end{array}$

    on $\Omega$ . If $u(x, t)$ is a solution of (2.8), then $u_{t}(x)=u(\mathcal{T}_{t}(x), t)$ is a solution of (2.9) and vice versa.We set

    (2.10) $\mathcal{A}’:=\frac{d}{dt}A(t)|_{t=0}$ , $\mathcal{A}’’:=\frac{d^{2}}{dt^{2}}A(t)|_{t=0}$ .

    Suppose that $\varphi 0\mathcal{T}_{t}$ has the following Taylor expansion:

    $\varphi\circ \mathcal{T}_{t}=\varphi+t\dot{\varphi}+\frac{1}{2}t^{2}\ddot{\varphi}+o(t^{2})$ .

    From the definition we find

    $\dot{\varphi}=S\cdot\nabla\varphi$ . $\ddot{\varphi}=T\cdot\nabla_{\forall^{\neg}}’+\mathcal{H}_{x}\varphi\cdot(S)^{2}$ .

    130

  • Let $u$ be a solution of (2.3). Then, we ($(differentiate$“ (2.9) and obtain the equation

    (2.11) $\{\begin{array}{l}(\nabla\dot{u}c, \nabla v)_{\Omega}=-(\mathcal{A}’\nabla u, \nabla c))_{\Omega},\forall v\in H_{0}^{1}(\Omega),\dot{u}_{\mathcal{L}}=(i on \partial\Omega.\end{array}$

    One more ($(differentiation$“ yields the equation

    (2.12)$(\nabla\ddot{u}_{\mathcal{L}}, \nabla v)_{\Omega}=-2(\mathcal{A}^{f}\nabla\dot{u}_{\mathcal{L}}, \nabla v)_{\Omega}-(\mathcal{A}’’\nabla u, \nabla v)_{\Omega}$, $\forall v\in H_{0}^{1}(\Omega)$ ,

    $\ddot{u}c=\ddot{\varphi}$ on $\partial\Omega$ .

    For solutions of these equations, we have the following lemma.

    Lemma 2.1 Suppose that $u,$ $u_{t},\dot{u}_{\mathcal{L}},\ddot{u}_{\mathcal{L}}\in H^{1}(\Omega)$ are solutions of the equations (2.3), (2.9),(2.11), (2.12), respectively. Then, $u_{t}$ has a Taylor expansion $u_{t}=u+t \dot{u}_{\mathcal{L}}+\frac{1}{2}t^{2}\ddot{u}c+o(t^{2})$ in$H^{1}(\Omega)$ . That is, the following is valid;

    $\lim_{tarrow 0+}\frac{\Vert\chi_{t}||_{H^{1}(\Omega)}}{t^{2}}=0$, $\chi_{t}:=v_{t}-(u+t\dot{u}_{\mathcal{L}}+\frac{1}{2}t^{2}ii_{\mathcal{L}})$ .

    Proof Since $S\in W^{1,\infty}(\tilde{\Omega};\mathbb{R}^{n})$ and $T\in L^{\infty}(\tilde{\Omega};\mathbb{R}^{n})$ , we see $A(t)\in L^{\infty}(\Omega;\mathbb{R}^{n^{2}})$ and

    (2.13) $\lim_{tarrow 0+}\frac{\Vert\alpha_{t}||_{L^{\infty}}}{t^{2}}=0$ , $\alpha_{t}:=A(t)-(1+t\mathcal{A}’+\frac{1}{2}t^{2}\mathcal{A}’’)$ .

    Define $z_{t},\dot{z},\ddot{z}$ as solutions of the following boundary value problems:

    $(\nabla z_{t}, \nabla v)_{\Omega}=0$ , $\forall v\in H_{0}^{1}(\Omega)$ , $z_{t}=\varphi 0\mathcal{T}_{t}$ on $\partial\Omega$ ,$(\nabla\dot{z}, \nabla v)_{\Omega}=0$ , $\forall v\in H_{0}^{1}(\Omega)$ , $\dot{z}=\dot{\varphi}$ on $\partial\Omega$ ,$(\nabla\ddot{z}, \nabla v)_{\Omega}=0$ , $\forall v\in H_{0}^{1}(\Omega)$ , $\sim\ddot{\vee}=\ddot{\varphi}$ on $\partial\Omega$ .

    Letting$\eta_{t}:=z_{t}-(u+t\dot{z}+\frac{1}{2}t_{\sim}^{2_{\vee}})$ , $\psi_{t}:=\varphi\circ \mathcal{T}_{t}-(\varphi+t\dot{\varphi}+\frac{1}{2}t^{2}\ddot{\varphi})$ ,

    we notice $\eta_{t}-\psi_{t}\in H_{0}^{1}(\Omega)$ . Since $0=(\nabla\eta_{t}, \nabla v)_{\Omega}$ for any $v\in H_{0}^{1}(\Omega)$ , we set $v$ $:=77t-\psi_{t}$ andobtain

    $\Vert\nabla\eta_{t}\Vert_{L^{2}(\Omega)}^{2}=(\nabla\eta_{t}, \nabla\eta_{t})_{\Omega}=(\nabla\eta_{t}, \nabla?l_{t}))_{\Omega}\leq\Vert\nabla_{7}h\Vert_{L^{2}(\Omega)}\Vert\nabla?i_{t}\Vert_{L^{2}(\Omega)}$ ,

    $\lim_{tarrow 0+}\frac{\Vert\nabla\eta_{t}||_{L^{2}(\Omega)}}{t^{2}}\leq\lim_{tarrow 0+}\frac{\Vert\nabla\psi_{t}||_{L^{2}(\Omega)}}{t^{2}}=0$ .

    Similarly, set$\beta_{t}$ $:=$ 娩一 $z_{t}-$ $(t$ $(\dot{u}c$ 一ゑ $)$ $+ \frac{1}{2}t^{2}(\ddot{u}_{C}-\sim\vee))\in H_{0}^{1}(\Omega)$

    Then, from (2.11), (2.12), we find that for any $v\in H_{0}^{1}(\Omega)$ ,

    $(A(t)\nabla\beta_{t}, \nabla v)_{\Omega}=((l-A(t))\nabla z_{t}, \nabla v)_{\Omega}$

    $-t( A(t)\nabla(\dot{u}c-\dot{z}), \nabla v)_{\Omega}-\frac{1}{2}t^{2}(A(t)\nabla(\ddot{u}_{\mathcal{L}}-$ を $), \nabla v)_{\Omega}$

    $=((I- A(t))\nabla(\eta_{t}+\frac{1}{2}t^{2}\ddot{u}c), \nabla v)_{\Omega}+t((I+tA’-A(t))\nabla\dot{u}c, \nabla v)_{\Omega}$

    $+$ $((I+t \mathcal{A}’+\frac{1}{2}t^{2}\mathcal{A}’’ - A(t))\nabla u, \nabla v)_{\Omega}$ .

    131

  • By (2.13) there exists a positive constant $\lambda$ such that, for any sufficiently small $t>0$ ,

    $\lambda\Vert\nabla v\Vert_{L^{2}(\Omega)}^{2}\leq(A(t)\nabla v, \nabla v)_{\Omega}$ , $\forall v\in H_{0}^{1}(\Omega)$ .

    Inserting $v=\beta_{t}$ into the above equation, we obtain

    $\frac{\lambda}{t^{2}}\Vert\nabla\beta_{t}\Vert_{L^{2}(\Omega)}\leq\Vert 1-A(t)\Vert_{L^{\infty}(\Omega)}(\frac{\Vert\nabla\eta_{t}||_{L^{2}(\Omega)}}{t^{2}}+\Vert\nabla..c\Vert_{L^{2}(\Omega)})$

    $+ \frac{1}{t}\Vert 1+t\mathcal{A}^{f}-A(t)\Vert_{L^{\infty}(\Omega)}\Vert\nabla\dot{u}_{\mathcal{L}}\Vert_{L^{2}(\Omega)}+\frac{1}{t^{2}}\Vert\alpha_{t}\Vert_{L^{\infty}(\Omega)}\Vert\nabla u\Vert_{L^{2}(\Omega)}$ .

    Therefore, we conclude $\lim_{tarrow 0+}\Vert\nabla\beta_{t}\Vert_{L^{2}(\Omega)}/t^{2}=0$ and complete the proof since $\chi_{t}=\beta_{t}+\eta_{t}$ . $\square$

    2.3 The Relationship between Eulerian and Lagrangian derivatives

    In this subsection we consider the relationship between Eulerian and Lagrangian derivatives.From (2.4) we immediately notice

    $\dot{u}_{\mathcal{L}}=\dot{u}_{\mathcal{E}}+S\cdot\nabla u$ in $\Omega$ .

    Since $(V\dot{u}_{\mathcal{L}}, \nabla v)=-(\mathcal{A}’\nabla u, \nabla v)$ and $(V\dot{u}\mathcal{E}, \nabla v)=0$ for any $v\in H_{0}^{1}(\Omega)$ , we have

    $(\nabla (S. \nabla u), \nabla v)_{\Omega}=-(\mathcal{A}’\nabla u, \nabla v)_{\Omega}$ , $\forall v\in H_{0}^{1}(\Omega)$ .

    Similarly, from (2.5) we obtain

    $\ddot{u}_{\mathcal{L}}=\ddot{u}_{\mathcal{E}}+2S\cdot\nabla\dot{u}_{\mathcal{E}}+T\cdot\nabla u+\mathcal{H}_{x}u$ . (S)2 in $\Omega$ .

    Since$(\nabla\ddot{u}_{\mathcal{L}}, \nabla v)_{\Omega}=-(2\mathcal{A}’\nabla\dot{u}_{\mathcal{L}}+\mathcal{A}’’\nabla u, \nabla v)_{\Omega}$, $(\nabla\ddot{u}_{\mathcal{E}}, \nabla v)=0$ , $\forall v\in H_{0}^{1}(\Omega)$ ,

    we have, for any $v\in H_{0}^{1}(\Omega)$ ,

    $(2\nabla(S\cdot Vu_{\mathcal{E}})+\nabla(T\cdot\nabla u)+\nabla(\mathcal{H}_{x}u\cdot(S)^{2}), \nabla v)_{\Omega}=-(2\mathcal{A}’\nabla\dot{u}_{\mathcal{L}}+\mathcal{A}’’\nabla u, \nabla v)_{\Omega}$ .

    2.4 Liouville’s Theorem

    In this section we prepare Liouville‘s theorem which plays an important role in calculus ofHadamard’s variation. Following Garabedian [1] and (2.2), we denote normal components of $S$and $T$ by $\delta\rho$ and $\delta^{2}\rho$ , respectively:

    (2.14) $\delta\rho$ $:=S\cdot\nu$ , $\delta^{2}\rho$ $:=T\cdot\nu=\nu^{t}DS(x)S(x)$ .

    Theorem 2.2 (Liouville’s Theorem) Let a sufficiently $srn$ooth function $c(x, t)$ be defined onthe domain $\Omega_{t}$ $:=\mathcal{T}_{t}(\Omega)$ for each $t\geq 0$ . Suppose also that $c(x, t),$ $c_{t}(x, t)$ $:= \frac{\partial c}{\partial t}(x, t)$ aremeasurable on $\Omega_{t}$ . Then, the following holds:

    (2.15) $\frac{d}{dt}(\int_{\Omega_{f}}c(x, t)dx)|_{t=0}=\int_{\Omega}(c_{t}(x, 0)+\nabla\cdot(c(x, 0)S(x)))dx$

    $= \int_{\Omega}c_{t}(x,0)dx+\langle c(\cdot,0),$ $\delta\rho\rangle_{\partial\Omega}$ .

    132

  • Proof. We may suppose without loss of generality that $\partial\Omega,$ $c,$ $c_{t}$ are all sufficiently smooth.The proof for general cases follows from the density property of $C^{\infty}(\tilde{\Omega})$ in $H^{1}(\tilde{\Omega})$ . Let $J\mathcal{T}_{t}(x)$be the Jacobi matrix of $\mathcal{T}_{t}(x)$ . Differentiating the both sides of

    (2.16) $\int_{\Omega_{f}}c(x, t)dx=\int_{\Omega}c(\mathcal{T}_{t}(x), t)\det(J\mathcal{T}_{t}(x))dx$

    with respect to $t$ , we have

    $\frac{d}{dt}(\int_{\Omega_{t}}c(x, t)dx)=\int_{\Omega}(c_{t}(\mathcal{T}_{t}(x), t)+\nabla c(\mathcal{T}_{t}(x), t)\cdot\frac{\partial}{\partial t}\mathcal{T}_{t}(x))\det(/\mathcal{T}_{t}(x))dx$

    (2.17)$+ \int_{\Omega}c(\mathcal{T}_{t}(x).t)\frac{\partial}{\partial t}\det(/\mathcal{T}_{t}(x))dx$ .

    Then, letting $tarrow 0+$ , we obtain the first equality of (2.15). Here, we use

    (2.18) $\frac{\partial}{\partial t}\det(/\mathcal{T}_{t}(x))|_{t=0}=\nabla\cdot S(x)$ .

    The second equality immediately follows from the divergence theorem. $\square$

    Corollary 2.3 Suppose that a function $f(x, t)$ is in $H^{1}(\Omega_{t})$ for each $t\geq 0$ and harmonic on $\Omega_{t}$with respect to $x\in \mathbb{R}^{n}$ . Then, we have

    $\frac{d}{dt}(\int_{\Omega_{t}}|\nabla_{x}f(x, t)|^{2}dx)t=0^{=2}\langle\frac{\partial f}{\partial\nu},\dot{f}_{\mathcal{E}}\rangle_{\partial\Omega}+\langle|\nabla f|^{2},$$\delta\rho\rangle_{\partial\Omega}$ .

    where $f(x):=f(x,0),\dot{f}_{\mathcal{E}}(x):=\partial\overline{t}\partial f(x, t)|_{t=0}$.

    Proof: Set $c(x, t);=|\nabla_{x}f(x, t)|^{2}$ and apply Theorem 2.2. $\square$

    We now try to obtain a second order Liouville’s theorem. Assume that $\partial\Omega,$ $c,$ $S$ are suffi-ciently smooth. We have obtained (2.17) by differentiating the both side of (2.16) with respectto $t$ . One more differentiation of the both side of (2.17) and letting $tarrow 0$ yield

    $\int_{\Omega}[c_{tt}(x, 0)+2\nabla_{x}c_{t}(x, 0)\cdot S+\nabla_{x}c(x, 0)\cdot T+\mathcal{H}_{x}c(x, 0) . (S)^{2}]dx$

    $+2 \int_{\Omega}[c_{t}(x,0)+\nabla_{x}c(x, 0)\cdot S](\nabla\cdot S)dx$

    $+ \int_{\Omega}c(x, 0)[\nabla\cdot T+2\sum_{i

  • We try to simplify this formula. Recall that $DS$ is the Jacobi matrix of the vector field S. Sinceit follows from the divergence theorem that

    2 $\int_{\Omega}(\nabla_{x}c(x, 0)\cdot S)(\nabla\cdot S)dx=2\int_{\partial\Omega}(\nabla_{x}c(x,0)\cdot S)\delta\rho ds$

    $-2 \int_{\Omega}\mathcal{H}_{x}c(x,0)\cdot(S)^{2}dx-2\int_{\Omega}((DS)S)\cdot\nabla_{x}c(x,0)dx$ ,

    we have

    $\frac{d^{2}}{dt^{2}}(\int_{\Omega},$ $c(x, t) dx)t=0=\int_{\Omega}c_{tt}(x, 0)dx+2\int_{\partial\Omega}c_{t}(x, 0)\delta\rho ds$

    $+ \int_{\partial\Omega}c(x,0)\delta^{2}\rho ds+2\int_{\partial\Omega}(\nabla_{x}c(x, 0)\cdot S)\delta\rho ds$

    $- \int_{\Omega}\mathcal{H}_{x}c(x,0)\cdot(S)^{2}dx-2\int_{\Omega}((DS)S)\cdot\nabla_{x}c(x, 0)dx$

    $+2 \int_{\Omega}c(x,0)\sum_{i

  • Therefore, we have

    $X+Y+Z=- \int_{\partial\Omega}(\nabla_{x}c\cdot S)\delta\rho+\sum_{i

  • Here, $\omega_{n}$ is the measure of $(n-1)$-dimensional sphere $S^{n-1}$ . Then, for sufficiently smoothfunction $f$ we have Green’s formula

    (3.1) $- \int_{\Omega}\Gamma(x-y)\Delta f(x)dx+\int_{\partial\Omega}\frac{\partial f}{\partial\nu}(x)\Gamma(x-y)ds_{x}=\int_{\partial\Omega}f(x)\frac{\partial}{\partial\nu_{x}}\Gamma(x-y)ds_{x}+f(y)$.

    For the fundamental solution $\Gamma(x-y)$ . define $u$ as a solution of the following boundary valueproblem:

    $\Delta u=0$ in $\Omega$ , $u(x)=-\Gamma(x-y)$ , $x\in\partial\Omega$ .Then,

    $G(x, y):=\Gamma(x-y)+u(x)$

    is the Green function of $\triangle$ on $\Omega$ . It follows from the definition that $G(x,y)=0$ for $x\in\partial\Omega$ and$y\in\Omega$ . Adding the following Green’s formula with respect to $f$ and $u$

    $- \int_{\Omega}u(x)\Delta f(x)dx+\int_{\partial\Omega}\frac{\partial f}{\partial\nu}(x)u(x)ds=\int_{\partial\Omega}f(x)\frac{\partial u}{\partial\nu}(x)ds$

    to (3.1), we obtain Green $s$ second formula

    (3.2) $f(y)=- \int_{\Omega}G(x,y)\Delta f(x)dx-\int_{\partial\Omega}f(x)\frac{\partial}{\partial\nu_{x}}G(x, y)ds_{x}$.

    3.1 First Variation

    Now, we consider domain perturbation $\Omega_{t}=\mathcal{T}_{t}(\Omega)$ defined in the previous section. The Greenfunction $G(x, y, t)$ on $\Omega_{t}$ is written as

    $G(x, y, t)=\Gamma(x-y)+u(x,t)$ ,

    where $u(x, t)$ is the harmonic function which satisfies

    (3.3) $\triangle_{x}u(x, t)=0$ in $\Omega_{t}$ , $u(x, t)=-\Gamma(x-y)$ , $x\in\partial\Omega_{t}$ .

    Obviously, we have $G(x, y, 0)=G(x, y)$ and $u(x,0)=u(x)$ . For two inner points $x,$ $y\in\Omega$ andsufficiently small $t>0$ , we have $x,$ $y\in\Omega_{t}$ . The first variation $\delta G(x, y)$ with respect to domainperturbation is defined by

    $\delta G(x, y):=\lim_{tarrow 0+}\frac{G(x,y,t)-G(x,y,0)}{t}=\lim_{tarrow 0+}\frac{u(x,t)-u(x,0)}{t}=\dot{u}_{\mathcal{E}}(x)$ ,

    and is equal to the Eulerian derivative $\dot{u}\mathcal{E}$ of $u$ .By (2.6), we confirm that $\dot{u}\mathcal{E}$ is a solution of the boundary value problem

    $\triangle\dot{u}\mathcal{E}=0$ in $\Omega$ ,

    $\dot{u}\mathcal{E}=S\cdot(-\nabla_{x}\Gamma(x-y)-\nabla u)=-S\cdot\nabla_{x}G(x,y)=-\delta\rho\frac{\partial}{\partial\nu_{x}}G(x,y)$ on $\partial\Omega$ .

    Here, we use the fact that $S\cdot\nabla_{x}G(x, y)=(S\cdot\nu)\frac{\partial}{\partial\nu_{x}}G(x, y)$ on $\partial\Omega$ . Applying the formula (3.2)to $\dot{u}_{\mathcal{E}}$ , we obtain Hadamard $s$ variational formula.

    Theorem 3.1 (Hadamard’s variational formula) The first variation. $\delta G(w, y)$ of the Greenfunction $G(w, y)$ of $\triangle$ with respect to domain perturbation is given by

    $\delta G(w,y)=\int_{\partial\Omega}\frac{\partial}{\partial\nu_{x}}G(x, y)\frac{\partial}{\partial\nu_{x}}G(x, w)\delta\rho ds_{x}$ , $\delta\rho:=S\cdot\nu$ .

    136

  • 3.2 Second VariationIn this subsection, we compute the second variation of the Green function with respect to domainperturbation. We prepare a lemma. Let a harmonic function $u(x, t)$ be a solution of the Dirichletproblem (3.3). Since $\delta G(x, y)=\dot{u}_{\mathcal{E}}(x)$ , we recall that

    (3.4) $\delta G(x, y)=-\delta\rho\frac{\partial}{\partial\nu_{x}}G(x, y)$ , $x\in\partial\Omega$ .

    Hence, for a harmonic function $g(x)$ , we find

    $0= \int_{\partial\Omega}(\delta G(x, y)+\delta\rho\frac{\partial}{\partial\nu_{x}}G(x, y))\frac{\partial g}{\partial\nu}(x)ds_{x}$

    $= \int_{\Omega}\nabla_{x}\delta G(x, y)\cdot\nabla g(x)dx+\int_{\partial\Omega}\delta\rho\frac{\partial}{\partial\nu_{x}}G(x, y)\frac{\partial g}{\partial\nu}(x)ds_{x}$ ,

    and obtain the following lemma.

    Lemma 3.2 For a harmonic function $g$ on $\Omega$ , the following equality holds:

    $\int_{\Omega}\nabla_{x}\delta G(x, y)\cdot\nabla g(x)dx=-\int_{\partial\Omega}\frac{\partial}{\partial\nu_{x}}G(x, y)\frac{\partial g}{\partial\nu}(x)\delta\rho ds_{x}$ .

    The second variation $\delta^{2}G(x, y)$ of the Green function $G(x, y)$ is defined by

    $\delta^{2}G(x, y):=\frac{\partial^{2}}{\partial t^{2}}G(x, y, t)|_{t=0}=\ddot{u}_{\mathcal{E}}(x)$ ,

    and, therefore, we only need to compute $\ddot{u}_{\mathcal{E}}$ . Recall that the harmonic function is a solution ofthe Dirichlet problem

    $\triangle u=0$ in $\Omega$ , $u=-\Gamma(\cdot-y)$ on $\partial\Omega$ .

    By (2.7), the boundary value of $\ddot{u}_{\mathcal{E}}$ on $\partial\Omega$ is

    $\ddot{u}\mathcal{E}=-2S\cdot\nabla_{x}\delta G(x, y)-T\cdot\nabla_{x}G(x, y)-\mathcal{H}_{x}G(x, y)\cdot(S)^{2}$ .

    Here, we use $G(x, y)=\Gamma(x-y)+u(x)$ and $\dot{u}_{\mathcal{E}}(x)=\delta G(x, y)$ . From (3.2), we find

    $\delta^{2}G(x, y)=\ddot{u}s(x)=2\int_{\partial\Omega}S\cdot\nabla_{w}\delta G(w, y)\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}$

    (3.5) $+ \int_{\partial\Omega}T\cdot\nabla_{w}G(w, y)\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}$

    $+ \int_{\partial\Omega}\mathcal{H}_{w}G(w, y)\cdot(S)^{2}\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}$ .

    We denote the first, second and third terms of the right-hand side of (3.5) by $X,$ $Y,$ $Z$ , respec-tively. As before, the term $Y$ can be written as

    (3.6) $Y= \int_{\partial\Omega}\frac{\partial}{\partial\nu_{w}}G(w, y)\frac{\partial}{\partial\nu_{x}}G(x, w)\delta^{2}\rho ds_{w}$ .

    To understand the terms $X$ and $Z$ , we consider the $(n-1)$-dimensional tangent space $\prime 1_{x}’\partial\Omega$of $\partial\Omega$ at $x\in\partial\Omega$ . Let $\{s_{1}, \cdots , s_{n-1}\}$ be the orthonormal basis of $\prime 1_{x}\partial\Omega$ . Then, $\{s_{1}, \cdots, s_{n-1}, \nu\}$

    137

  • is an orthonormal basis of the tangent space $\prime 1_{x}\mathbb{R}^{n}$ at $x\in \mathbb{R}^{n}$ . For a generic function $f$ , directionalderivatives are defined by

    $\frac{\partial f}{\partial\nu}=\nabla f\cdot\nu$ , $\frac{\partial f}{\partial s_{i}}=\nabla f\cdot s_{i}$ , $i=1,$ $\cdots,$ $n-1$ .

    Thus, defining the orthogonal matrix $P$ by $P:=(s_{1}\ldots., s_{n-1}, \nu)$ , we may write

    $( \frac{\partial f}{\partial s_{1}},$ $\cdot\cdot\cdot$ $\frac{\partial f}{\partial s_{n-1}},$ $\frac{\partial f}{\partial\nu})=(\nabla f)P$, or

    (3.7) $( \nabla f)^{T}=\sum_{i=1}^{n-1}s_{i}\frac{\partial f}{\partial s_{i}}+\nu\frac{\partial f}{\partial\nu}$ and $\nabla=\sum_{i=1}^{n-1}s_{i}^{T}\frac{\partial}{\partial s_{i}}+\nu^{T}\frac{\partial}{\partial\nu}$.

    If we write $S$ as

    (3.8) $S=\sum_{i=1}^{n-1}\mu_{i}s_{i}+\delta\rho\nu$ , $\delta\rho=S\cdot\nu$ , $\mu_{i}=S$ . si, $i=1,$ $\cdots,$ $7t-1$

    on $\partial\Omega$ , we obtain

    $S\cdot\nabla_{w}\delta G(w, y)=\sum_{i=1}^{n-1}\mu_{i}\frac{\partial}{\partial s_{i}}\delta G(w, y)+\delta\rho\frac{\partial}{\partial\nu_{w}}\delta G(w, y)$.

    Using Lemma 3.2 with $g:=\delta G$ , the term $X$ (the first term of the right-hand side of (3.5)) iswritten as

    $X=2 \sum_{i=1}^{n-1}\int_{\partial\Omega}\mu_{i}\frac{\partial}{\partial s_{i}}\delta G(w,y)\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}+2\int_{\partial\Omega}\delta\rho\frac{\partial}{\partial\nu_{w}}\delta G(w,y)\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}$

    $=2 \sum_{i=1}^{n-1}\int_{\partial\Omega}\mu_{i}\frac{\partial}{\partial s_{i}}\delta G(w,y)\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}-2\int_{\Omega}\nabla_{w}\delta G(w, y)\cdot\nabla_{x}\delta G(x, w)ds_{w}$.

    Next, we try to rewrite the third term $Z$ of the right-hand side of (3.5). To this end, weconsider the curved coordinate defined by $\{s_{1}, \cdots, s_{n-1}, \nu\}$ in the neighborhood of $x\in\partial\Omega$ andsecond order differentiation on the coordinate. For a $C^{2}$ class generic function $f$ , the Hessematrix $\mathcal{H}f$ is written by $\mathcal{H}f=\nabla(\nabla f)^{T}$ and

    $\mathcal{H}f=\nabla(\sum_{i=1}^{n-1}s_{i}\frac{\partial f}{\partial s_{i}}+\nu\frac{\partial f}{\partial\nu})$

    $= \sum_{i=1}^{n-1}\frac{\partial f}{\partial s_{i}}Ds_{i}+\frac{\partial f}{\partial\nu}D\nu+\sum_{i=1}^{n-1}s_{i}\nabla(\frac{\partial f}{\partial s_{i}})+\nu\nabla(\frac{\partial f}{\partial\nu}I$

    (3.9) $= \sum_{i=1}^{n-1}\frac{\partial f}{\partial s_{i}}Ds_{i}+\frac{\partial f}{\partial\nu}D\nu+\sum_{i,j=1}^{n-1}\frac{\partial^{2}f}{\partial s_{j}\partial s_{i}}s_{i}s_{j}^{T}+\sum_{i=1}^{n-1}\frac{\partial^{2}f}{\partial s_{i}\partial\nu}(s_{i}\nu^{T}+\nu s_{i}^{T})+\frac{\partial^{2}f}{\partial\nu^{2}}\nu\nu^{T}$.

    138

  • Similarly, $\triangle f$ is $co\dot{r}nputed$ as

    $\triangle f=\nabla\cdot\nabla f=\nabla\cdot(\sum_{i=1}^{n-1}s_{i}\frac{\partial f}{\partial s_{i}}+\nu\frac{\partial f}{\partial\nu})$

    $=( \nabla\cdot\nu)\frac{\partial f}{\partial\nu}+\nu\cdot\nabla\frac{\partial f}{\partial\nu}+\sum_{i=1}^{n-1}((\nabla\cdot s_{i})\frac{\partial f}{\partial s_{i}}+s_{i}\cdot\nabla\frac{\partial f}{\partial s_{i}})$

    (3.10) $=( \nabla\cdot\nu)\frac{\partial f}{\partial\nu}+\frac{\partial^{2}f}{\partial\nu^{2}}+\sum_{i=1}^{n-1}((\nabla\cdot s_{i})\frac{\partial f}{\partial s_{i}}+\frac{\partial^{2}f}{\partial s_{i}^{2}})$ .

    Since the Green function $G$ satisfies $G(x, y)=0$ on $\partial\Omega$ , we have

    $\frac{\partial}{\partial s_{i}}G(x, y)=\frac{\partial^{2}}{\partial s_{i}\partial s_{j}}G(x, y)=0$, $i,j=1,$ $\cdots,$ $?-1$ .

    and, thus,

    $0= \triangle_{w}G(w, y)=(\nabla\cdot\nu)\frac{\partial}{\partial\nu_{w}}G(w, y)+\frac{\partial^{2}}{\partial\nu_{w}^{2}}G(w, y)$ , $w\in\partial\Omega$ .

    Applying these results to computation of $\mathcal{H}_{w}G(w, y)\cdot S^{2}$ with $S=\sum_{i=1}^{n-1}\mu_{i}s_{i}+\delta\rho\nu$ , we obtain

    $\mathcal{H}_{w}G(w, y)\cdot(S)^{2}=2\delta\rho\sum_{i=1}^{n-1}\mu_{i}\frac{\partial^{2}}{\partial s_{i}\partial\nu_{w}}G(w,y)-(\nabla\cdot\nu)(\delta\rho)^{2}\frac{\partial}{\partial\nu_{w}}G(w, y)$

    $+( \sum_{i,j=1}^{n-1}\mu_{i}\mu_{j}s_{i}^{T}(D\nu)s_{j}+2\sum_{i=1}^{n-1}\mu_{i}\delta ps_{i}^{T}(D\nu)\nu+(\delta\rho)^{2}\nu^{T}(D\nu)\nu)\frac{\partial}{\partial\nu_{w}}G(w, y)$

    $=2 \delta\rho\sum_{i=1}^{n-1}\mu_{i}\frac{\partial^{2}}{\partial s_{i}\partial\nu_{w}}G(w, y)+(\sum_{i=1}^{n-1}\kappa_{i}(\mu_{i}^{2}-(\delta\rho)^{2}))\frac{\partial}{\partial\nu_{w}}G(w, y)$.

    Here, we use the fact $\nabla\nu=\sum_{i=1}^{n-1}\kappa_{i}s_{i}s_{i}^{t},$ $\nabla\cdot\nu=$ tr $( \nabla\nu)=\sum_{i=1}^{n-1}\kappa_{i}$ , where $\kappa_{i}$ is the curvatureof the cross-section of $(n-1)$-dimensional surface $\partial\Omega$ by a two-dimensional plane defined by $s_{i}$and $\nu$ . Therefore, the third term $Z$ of the right-hand side of (3.5) is written by

    $Z= \int_{\partial\Omega}\sum_{i=1}^{n-1}\kappa_{i}(\mu_{i}^{2}-(\delta\rho)^{2})\frac{\partial}{\partial\nu_{w}}G(w, y)\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}$

    $+2 \int_{\partial\Omega}\sum_{i=1}^{n-1}\mu_{i}\delta\rho\frac{\partial^{2}}{\partial s_{i}\partial\nu_{w}}G(w.y)\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}$ .

    Noticing (3.4), we see the equality

    $\frac{\partial}{\partial s_{i}}\delta G(w, y)+\delta\rho\frac{\partial^{2}}{\partial s_{i}\partial\nu_{w}}G(w, y)=-\frac{\partial(\delta\rho)}{\partial s_{i}}\frac{\partial}{\partial\nu_{w}}G(w, y)$.

    Computing $X+Z$ using this equality, we find

    $X+Z=-2 \int_{\Omega}\nabla_{w}\delta G(w, y)\nabla_{w}\delta G(x, w)ds_{w}$

    (3.11)$+ \int_{\partial\Omega}\sum_{i=1}^{n-1}[\kappa_{i}(\mu_{i}^{2}-(\delta\rho)^{2})-2\mu_{i}\frac{\partial(\delta\rho)}{\partial s_{i}}]\frac{\partial}{\partial\nu_{w}}G(w, y)\frac{\partial}{\partial\nu_{x}}G(x.e\iota)ds_{w}$ .

    139

  • Gathering (3.5), (3.6), and (3.11) we obtain

    $\delta^{2}G(x, y)=-2\int_{\Omega}\nabla_{w}\delta G(w, y)\nabla_{w}\delta G(x, w)ds_{w}+\int_{\partial\Omega}\chi\frac{\partial}{\partial\nu_{w}}G(w, y)\frac{\partial}{\partial\nu_{x}}G(x, w)ds_{w}$,

    $\chi:=\delta^{2}\rho+\sum_{i=1}^{n-1}[\kappa_{i}(\mu_{i}^{2}-(\delta\rho)^{2})-2\mu_{i}\frac{\partial(\delta\rho)}{\partial s_{i}}]$ .

    We further try to simplify $\chi$ . At first, since $\partial\partial t=\kappa_{i}s_{i}$ , we notice

    $\sum_{i=1}^{n-1}\kappa_{i}\mu_{i}^{2}=\sum_{i=1}^{n-1}\kappa_{i}(S\cdot s_{i})^{2}=\sum_{i=1}^{n-1}S\cdot s_{i}(\frac{\partial(\delta\rho)}{\partial s_{i}}-$ $( \frac{\partial S}{\partial s_{i}}$ . $\nu))$ .

    Thus, recalling from (3.7) that $D S=\sum_{n=1\mathcal{T}s_{i}}^{n-1\partial S}s_{i}^{T}+\frac{\partial S}{\partial\nu}\nu^{T}$ , we see

    $\sum_{i=1}^{n-1}S\cdot s_{i}(\frac{\partial S}{\partial s_{i}}\cdot\nu)=\sum_{i=1}^{n-1}S\cdot s_{i}(\nu^{T}(DS)s_{i})=\nu^{T}(DS)(\sum_{i=1}^{n-1}(S\cdot s_{i})s_{i})$

    (3.12)

    $= \nu^{T}(DS)(S-\delta\rho\nu)=\delta^{2}\rho-\delta\rho(\frac{\partial S}{\partial\nu}\cdot\nu)=\delta^{2}\rho-\frac{1}{2}\frac{\partial(\delta^{-}\rho)^{2}}{\partial\nu}$ .

    Here, we use the facts $\delta^{2}\rho=\nu^{T}(DS)S$ by (2.14) and $\frac{\partial\nu}{\partial\nu}=0$ . Similarly, since $S\cdot\nabla=\sum_{n=1}^{n-1}(S\cdot$$s_{i})_{Ts_{i}}^{\partial}+(S\cdot\nu)_{T\nu}^{\partial}$ by (3.7), we have

    (3.13) $\sum_{i=1}^{n-1}S\cdot s_{i}\frac{\partial(\delta\rho)}{\partial s_{i}}=(S\cdot\nabla)\delta\rho-\delta\rho\frac{\partial(\delta\rho)}{\partial\nu}=(S\cdot\nabla)\delta\rho-\frac{1}{2}\frac{\partial(\delta\rho)^{2}}{\partial\nu}$ .

    Letting $\tilde{\kappa}$ $:= \sum_{i=1}^{n-1}\kappa_{i},$ $\chi$ is rewritten as

    $\chi=-\tilde{\kappa}(\delta\rho)^{2}-(S\cdot\nabla)\delta p+\frac{\partial(\delta^{-}\rho)^{2}}{\partial\nu}$ .

    If the domain perturbation $\mathcal{T}_{t}(x)$ satisfies $S\cdot s_{i}=0,$ $i=1,$ $\cdots,$ $n-1$ , it follows from (3.12),(3.13) that

    (3.14) $\delta^{2}\rho=(S\cdot\nabla)\delta\rho=\frac{1}{2}\frac{\partial(\delta\rho)^{2}}{\partial\nu}$ .

    Therefore, in this case, we find$\chi=-\tilde{\kappa}(\delta\rho)^{2}+\delta^{2}\rho$.

    So far, computation has been done under smoothness assumptions. A usual density argu-ment yields the following theorem:

    Theorem 3.3 (Hadamard’s second variational formula) Let $\Omega\subset \mathbb{R}^{n}$ be a Lipschitz do-main and $\mathcal{T}_{t}(x)$ be a $W^{1,\infty}$ class domain perturbation on $\Omega$ . Then, the second variation $\delta^{2}G(w, y)$of the Green function $G(w, y)$ of Laplacian on $\Omega$ is written by

    $\delta^{2}G(x, y)=\langle\chi\frac{\partial}{\partial\nu}G(x, \cdot),$ $\frac{\partial}{\partial\nu}G(\cdot, y)\rangle_{\partial\Omega}-2(\nabla\delta G(x, \cdot), \nabla\delta G(\cdot, y))_{\Omega}$ ,

    $\chi=-\tilde{\kappa}(\delta p)^{2}-(S\cdot\nabla)\delta\rho+\frac{\partial(\grave{\delta}\rho)^{2}}{\partial\nu}$ , $\delta\rho:=S\cdot\nu$ , $\tilde{\kappa}:=\sum_{i=1}^{n-1}t\mathfrak{i}_{i}$ ,

    140

  • where $\{s_{i}\}_{i=1}^{n-1}$ is an orthonormal basis of the tangent space of $\partial\Omega$ and $\kappa_{i}$ is the curvature of $\partial\Omega$along si. In particular, if the perturbation satisfies S. $s_{i}=0,$ $i=1,$ $\cdots,$ $n-1$ , we have

    (3.15) $\delta^{2}G(x, y)=\langle(\delta^{2}\rho-\tilde{\kappa}(\delta\rho)^{2})\frac{\partial}{\partial\nu}G(x, \cdot),$ $\frac{\partial}{\partial\nu}G(\cdot, y)\rangle_{\partial\Omega}-2(\nabla\delta G(x, \cdot), \nabla\delta G(\cdot, y))_{\Omega}$ .

    Remark: Garabedian and Schiffer [2] dealt with domain perturbation such as

    $\mathcal{T}_{t}(x)=x+th(x)\nu(x)$ , $t\geq 0$ ,

    where $h(x)$ is a scalar function defined on $\partial\Omega$ . In this case, Hadamard’s second variationalformula is (3.15) with $\delta\rho=h$ and $\delta^{2}\rho(x)=0$ which is exactly same to Garabedian-Schiffer’sformula [2]. Therefore, Theorem 3.3 is an extension of Garabedian-Schiffer’s formula.

    References

    [1] P.R. GARABEDIAN, Partial Differential Equations (2nd ed.) (1986) Chelsea.

    [2] P.R. GARABEDIAN, M. SCHIFFER, Convexity of domain functionals, J. Anal. Math., 2(1952-53), 281-368.

    [3] J. HADAMARD, M\’emoire sur le probl\‘eme $d$ ‘analyse relatif \‘a 1‘\’equilibre des plaques \’elastiquesencastrees. M\’emoires present\’es par divers savants \‘a 1‘Acad\’emie des Sciences, Vol. 33 (1908),1-128 (Oeuvres., 2 (1968), 515-631).

    [4] M. SCHIFFER, Hadamard $s$ formula and variation of domain-functions, Amer. J. Math., 68(1946), 417-448.

    141